[External] Re: GHC memory usage when typechecking from source vs. loading ModIfaces
Erdi, Gergo
Gergo.Erdi at sc.com
Tue Apr 1 12:24:14 UTC 2025
PUBLIC
OK so this `ModuleGraph` leak has the potential to be a huge explainer because of the large number of ModuleNodes and NodeKey_Modules in the heap profile. It turns out the reason we had one ModuleGraph per module was because I was populating the ModuleGraph by repeated calls to `extendMG`, where the new edges were created via `mgModSummaries` on the old `ModuleGraph`, and so the new edge list was a thunk pointing to the old `ModuleGraph`.
I’ve now changed it to a single `mkModuleGraph` call, which gets rid of the extra ModuleGraphs (see attached profile), but unfortunately, the ModuleNodes/ edge NodeKey_Modules are still not shared. Here’s how I build the graph, trying hard with a Data.Map.Map to ensure only use one NodeKey_Module per each target module:
```
registerModulesToGraph :: GhcMonad m => [(UnitId, ModuleName, [ModuleName])] -> m ()
registerModulesToGraph entries = modifySessionM \hsc_env -> do
mod_deps <- for entries \(unit, mod_name, deps) -> do
mod <- liftIO $ registerToModuleFinder hsc_env unit mod_name
let ms = dummyModSummary (hsc_dflags hsc_env) mod deps fingerprint0
pure (ms, deps)
let nodes = [ModuleNode (edges deps) ms | (ms, deps) <- mod_deps]
pure hsc_env{ hsc_mod_graph = mkModuleGraph nodes }
where
edge_map :: Map.Map ModuleName NodeKey
edge_map = Map.fromList [(mod_name, NodeKey_Module $ ModNodeKeyWithUid (notBoot mod_name) uid) | (uid, mod_name, _) <- entries]
edges deps = map (edge_map Map.!) deps
registerToModuleFinder :: HscEnv -> UnitId -> ModuleName -> IO Module
registerToModuleFinder hsc_env unitId modName = do
addHomeModuleToFinder (hsc_FC hsc_env) homeUnit (notBoot modName) location
where
homeUnit = DefiniteHomeUnit unitId Nothing
location = dummyModLocation modName
```
But, as you can see in the attached profile, I still get way more than 2k ModuleNodes and NodeKey_Modules….
From: Matthew Pickering <matthewtpickering at gmail.com>
Sent: Tuesday, April 1, 2025 5:44 PM
To: Erdi, Gergo <Gergo.Erdi at sc.com>
Cc: GHC Devs <ghc-devs at haskell.org>; ÉRDI Gergő <gergo at erdi.hu>; Montelatici, Raphael Laurent <Raphael.Montelatici at sc.com>; Dijkstra, Atze <Atze.Dijkstra at sc.com>
Subject: [External] Re: GHC memory usage when typechecking from source vs. loading ModIfaces
Hi Gergo,
I looked in the detailed pane, searched for ModuleGraph, hovered my mouse over the "ModuleGraph` constructor, recorded the number of live bytes, divided that by 32.
Matt
On Tue, Apr 1, 2025 at 7:04 AM Erdi, Gergo <Gergo.Erdi at sc.com<mailto:Gergo.Erdi at sc.com>> wrote:
PUBLIC
OK scratch that, I was looking at wrong ghc-debug output. Indeed there are 2301 ModuleGraphs in the heap at the end of typechecking :O
From: Erdi, Gergo
Sent: Tuesday, April 1, 2025 12:51 PM
To: Matthew Pickering <matthewtpickering at gmail.com<mailto:matthewtpickering at gmail.com>>
Cc: GHC Devs <ghc-devs at haskell.org<mailto:ghc-devs at haskell.org>>; ÉRDI Gergő <gergo at erdi.hu<mailto:gergo at erdi.hu>>; Montelatici, Raphael Laurent <Raphael.Montelatici at sc.com<mailto:Raphael.Montelatici at sc.com>>; Dijkstra, Atze <Atze.Dijkstra at sc.com<mailto:Atze.Dijkstra at sc.com>>
Subject: Re: GHC memory usage when typechecking from source vs. loading ModIfaces
This sounds extremely interesting, but I don’t understand where you are getting this number from! How do you see in the eventlog HTMLs that I’ve included that there are ~2000 ModuleGraphs? I’ve now tried using ghc-debug to find all ModuleGraph constructors at two points in the run: just before typechecking the first module (after all the extendMG calls) and just after typechecking the last module, and even in the cold case I only see 1 ModuleGraph before and 13 ModuleGraphs after.
Also, what do you mean by “precisely one loaded per interface loaded into the EPS”? Since my repro has 2294 modules, wouldn’t that mean 2294 ModuleGraphs by that metric?
From: Matthew Pickering <matthewtpickering at gmail.com<mailto:matthewtpickering at gmail.com>>
Sent: Saturday, March 29, 2025 1:53 AM
To: Erdi, Gergo <Gergo.Erdi at sc.com<mailto:Gergo.Erdi at sc.com>>
Cc: GHC Devs <ghc-devs at haskell.org<mailto:ghc-devs at haskell.org>>; ÉRDI Gergő <gergo at erdi.hu<mailto:gergo at erdi.hu>>; Montelatici, Raphael Laurent <Raphael.Montelatici at sc.com<mailto:Raphael.Montelatici at sc.com>>; Dijkstra, Atze <Atze.Dijkstra at sc.com<mailto:Atze.Dijkstra at sc.com>>
Subject: [External] Re: GHC memory usage when typechecking from source vs. loading ModIfaces
Hi Gergo,
I quickly tried building `Cabal` with the master branch. There is precisely 1 ModuleGraph allocated for the home session, and precisely one loaded per interface loaded into the EPS. No leaky behaviour like you can see in your eventlogs.
It seems there are about 2000 live module graphs in your program, are you doing something with the API to create this many?
Cheers,
Matt
On Fri, Mar 28, 2025 at 12:40 PM Matthew Pickering <matthewtpickering at gmail.com<mailto:matthewtpickering at gmail.com>> wrote:
HI Gergo,
Do you have a (synthetic?) reproducer? You have probably identified some memory leak. However, without any means to reproduce it becomes very difficult to investigate. I feel like we are getting into very precise details now, where speculating is not going to be so useful.
It seems like this is an important thing for you and your company. Is there any budget to pay for some investigation? If that was the case then some effort could be made to create a synthetic producer and make the situation more robust going into the future if your requirements were precisely understood.
Cheers,
Matt
On Fri, Mar 28, 2025 at 10:12 AM Erdi, Gergo <Gergo.Erdi at sc.com<mailto:Gergo.Erdi at sc.com>> wrote:
PUBLIC
Just to add that I get the same "equalizing" behaviour (but in a more "natural" way) if instead of deepseq-ing the ModuleGraph upfront, I just call `hugInstancesBelow` before processing each module. So that's definitely one source of extra memory usage. I wonder if it would be possible to rebuild the ModuleGraph periodically (similar to the ModDetails dehydration), or if there are references to it stored all over the place from `HscEnv`s scattered around in closures etc. (basically the same problem the HPT had before it was made into a mutable reference).
-----Original Message-----
From: ghc-devs <ghc-devs-bounces at haskell.org<mailto:ghc-devs-bounces at haskell.org>> On Behalf Of Erdi, Gergo via ghc-devs
Sent: Friday, March 28, 2025 4:49 PM
To: Matthew Pickering <matthewtpickering at gmail.com<mailto:matthewtpickering at gmail.com>>; GHC Devs <ghc-devs at haskell.org<mailto:ghc-devs at haskell.org>>
Cc: ÉRDI Gergő <gergo at erdi.hu<mailto:gergo at erdi.hu>>; Montelatici, Raphael Laurent <Raphael.Montelatici at sc.com<mailto:Raphael.Montelatici at sc.com>>; Dijkstra, Atze <Atze.Dijkstra at sc.com<mailto:Atze.Dijkstra at sc.com>>
Subject: [External] Re: GHC memory usage when typechecking from source vs. loading ModIfaces
Hi,
Unfortunately, I am forced to return to this problem. Everything below is now in the context of GHC 9.12 plus the mutable HPT patch backported.
My test case is typechecking a tree of 2294 modules that form the transitive closure of a single module's dependencies, all in a single process. I have done this typechecking three times, here's what `+RTS -s -RTS` reports for max residency:
* "cold": With no on-disk `ModIface` files, i.e. from scratch: 537 MB
* "cold-top": With all `ModIface`s already on disk, except for the
single top-level module: 302 MB
* "warm": With all `ModIface`s already on disk: 211 MB
So my stupidly naive question is, why is the "cold" case also not 302 MB?
In earlier discussion, `ModDetails` unfolding has come up. Dehydrating `ModDetails` in the HPT all the time is disastrous for runtime, but based on this model I would expect to see improvements from dehydrating "every now and then". So I tried a stupid simple example where after every 100th typechecked module, I run this function on the topologically sorted list of modules processed so far:
```
dehydrateHpt :: HscEnv -> [ModuleName] -> IO () dehydrateHpt hsc_env mods = do
let HPT{ table = hptr } = hsc_HPT hsc_env
hpt <- readIORef hptr
for_ mods \mod -> for_ (lookupUDFM hpt mod) \(HomeModInfo iface _details _linkable) -> do
!details <- initModDetails hsc_env iface
pure ()
```
Buuut the max residency is still 534 MB (see "cold-dehydrate"); in fact, the profile looks exactly the same.
Speaking of the profile, in the "cold" case I see a lot of steadily increasing heap usage from the `ModuleGraph`. I could see this happening if typechecking from scratch involves more `modulesUnder` calls which in turn force more and more of the `ModuleGraph`. If so, then maybe this could be worked around by repeatedly remaking the `ModuleGraph` just like I remake the `ModDetails` above. I tried getting rid of this effect by `deepseq`'ing the `ModuleGraph` at the start, with the idea being that this should "equalize" the three scenarios if this really is a substantial source of extra memory usage. This pushes up the warm case's memory usage to 381 MB, which is promising, but I still see a `Word64Map` that is steadily increasing in the "cold-force-modulegraph" case and contributes a lot to the memory usage. Unfortunately, I don't know where that `Word64Map` is (it could be any `Unique`-keyed environment...).
So I am now stuck at this point. To spell out my goal explicitly, I would like to typecheck one module after another and not keep anything more in memory around than if I loaded them from `ModIface` files.
Thanks,
Gergo
p.s.: I couldn't find a way in the EventLog output HTML to turn event markers on/off or filter them, so to avoid covering the whole graph with gray lines, I mark only every 100th module.
From: Matthew Pickering <matthewtpickering at gmail.com<mailto:matthewtpickering at gmail.com>>
Sent: Wednesday, February 12, 2025 7:08 PM
To: ÉRDI Gergő <gergo at erdi.hu<mailto:gergo at erdi.hu>>
Cc: Erdi, Gergo <Gergo.Erdi at sc.com<mailto:Gergo.Erdi at sc.com>>; Zubin Duggal <zubin at well-typed.com<mailto:zubin at well-typed.com>>; Montelatici, Raphael Laurent <Raphael.Montelatici at sc.com<mailto:Raphael.Montelatici at sc.com>>; GHC Devs <ghc-devs at haskell.org<mailto:ghc-devs at haskell.org>>
Subject: [External] Re: GHC memory usage when typechecking from source vs. loading ModIfaces
You do also raise a good point about rehydration costs.
In oneshot mode, you are basically rehydrating the entire transitive closure of each module when you compile it, which obviously results in a large amount of repeated work. This is why people are investigating ideas of a persistent worker to at least avoid rehydrating all external dependencies as well.
On Mon, Feb 10, 2025 at 12:13 PM Matthew Pickering <mailto:matthewtpickering at gmail.com<mailto:matthewtpickering at gmail.com>> wrote:
Sure, you can remove them once you are sure they are not used anymore.
For clients like `GHCi` that doesn't work obviously as they can be used at any point in the future but for a batch compiler it would be fine.
On Mon, Feb 10, 2025 at 11:56 AM ÉRDI Gergő <mailto:gergo at erdi.hu<mailto:gergo at erdi.hu>> wrote:
On Mon, 10 Feb 2025, Matthew Pickering wrote:
> I wonder if you have got your condition the wrong way around.
>
> The only "safe" time to perform rehydration is AFTER the point it can
> never be used again.
>
> If you rehydrate it just before it is used then you will repeat work
> which has already been done. If you do this, you will always have a
> trade-off between space used and runtime.
Oops. Yes, I have misunderstood the idea. I thought the idea was that after loading a given module into the HPT, its ModDetails would start out small (because of laziness) and then keep growing in size as more and more of it are traversed, and thus forced, during the typechecking of its dependees, so at some point we would want to reset that into the small initial representation as created by initModDetails.
But if the idea is that I should rehydrate modules when they can't be used anymore, then that brings up the question why even do that, instead of straight removing the HomeModInfos from the HPT?
----------------------------------------------------------------------
----------------------------------------------------------------------
This email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please delete all copies and notify the sender immediately. You may wish to refer to the incorporation details of Standard Chartered PLC, Standard Chartered Bank and their subsidiaries together with Standard Chartered Bank’s Privacy Policy via our public website.
----------------------------------------------------------------------
This email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please delete all copies and notify the sender immediately. You may wish to refer to the incorporation details of Standard Chartered PLC, Standard Chartered Bank and their subsidiaries together with Standard Chartered Bank’s Privacy Policy via our main Standard Chartered PLC (UK) website at sc. com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-devs/attachments/20250401/c3387938/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: cold-no-modulegraph-leak.zip
Type: application/x-zip-compressed
Size: 1337463 bytes
Desc: cold-no-modulegraph-leak.zip
URL: <http://mail.haskell.org/pipermail/ghc-devs/attachments/20250401/c3387938/attachment-0001.bin>
More information about the ghc-devs
mailing list